|
Abstract:
|
1-in-7 electronically monitored patients in the stepdown ward (a high-acuity hospital environment) suffer cardiac arrest or readmission to the ICU under emergency conditions. Advanced detection of patient deterioration could prevent many of these adverse events. The current state-of-the-art relies on heuristic threshold-based "Early Warning Scores" (which are often manual), or kernel density estimates of a representative population's vital-sign data. Both methods assume an i.i.d. relationship between vital-sign data and can only assess a patient's present risk. They are unable to forecast future risk. Gaussian Process (GP) regression addresses both the i.i.d. assumption and forecasting: dependencies between vital-sign data within a time-series are modelled by the GP, allowing explicit modelling of vital-sign dynamics; the plausible range of future values is represented as a distribution over a function of time. The posterior GP provides a stochastic representation of the patient's physiological time-series, with desirable characteristics, such as the latent mean function, and the distribution of noisy observations around the mean. The posterior fits of different time-series are amenable to comparison, for example, with information theoretic measures. A dictionary of clinically validated healthy patients can then be used as a reference point for new unclassified patients. Furthermore, a patient's current values can be compared to those previously forecast, to deduce whether a step-change in physiology has occurred. These methods become more robust when model parameters are regularized using priors elicited from physiological phenomena.
|
